Higher-Level Cognition - Computational Explorations in Cognitive Neuroscience - page 377

There also appear to be some functional differences
associated with the dorsolateral and ventromedial re-
gions of prefrontal cortex, but the precise characteri-
zation of these differences is controversial. For exam-
ple, some theories hold that dorsolateral areas are im-
portant for active/working memory, while ventromedial
areas perform behavioral inhibition and other kinds of
more affective processing (e.g., Fuster, 1989; Diamond,
1990). Others hold that the dorsolateral region is im-
portant for more complex processing, while the ven-
tromedial region performs simpler memory functions
(e.g., Petrides, 1996). In the context of the dynamic
categorization model explored below, we posit a dorso-
lateral/ventromedial distinction between more abstract
and more concrete representations, respectively. How-
ever, we regard all of these ideas as speculative and in-
complete at this point, because there is simply insuffi-
cient evidence.
At a finer grain of organization, we reviewed in chap-
ter 9 that there is evidence that the prefrontal cortex may
have more isolated patterns of connectivity — neurons
there appear to be interconnected within self-contained
“stripe” patterns (Levitt et al., 1993), and iso-coding
microcolumns of neurons have been recorded (Rao
et al., 1999). This biological data can influence ideas
about the nature of frontal representations — isolated
stripes may facilitate the development of representa-
tions that are more easily and flexibly combined with
each other for novel task performance (O'Reilly et al.,
1999a; O'Reilly et al., 1999b).
One other important aspect of frontal biology has to
do with its extensive interconnectivity with the basal
ganglia. The frontal cortex provides one of the primary
inputs to the striatum of the basal ganglia, and the basal
ganglia project via the thalamus back up to the frontal
cortex, creating a series of “loops” (Alexander, De-
Long, & Strick, 1986). As discussed in chapters 7 and
9, these loops through the basal ganglia may provide
a fine-grained gating/motor initiation mechanism. This
intimate association of the basal ganglia and the frontal
cortex can help to explain why basal ganglia lesions
(e.g., in Parkinson's disease) often produce frontal-like
behavioral deficits (which are reviewed in a later sec-
tion).

11.3 Controlled Processing and the Stroop Task

Our first exploration focuses on the distinction between
controlled and automatic processing using a model of
the Stroop task (Stroop, 1935). As mentioned in the
introduction, the controlled-processing aspect of the
Stroop task involves naming the ink color of a conflict-
ing color word (e.g., saying “red” to the word “green”
printed in red ink). Our model is based on that devel-
oped by Cohen and colleagues (Cohen et al., 1990; Co-
hen & Servan-Schreiber, 1992; Cohen & Huston, 1994),
and demonstrates how top-down activation-based bias-
ing from the frontal cortex can enable controlled pro-
cessing by overriding prepotent associations encoded
in the posterior cortex. That is, the frontal activation
supports the weaker process of color naming over the
prepotent response of word reading. Frontal cortex is
uniquely important for this biasing function because it
can robustly maintain the appropriate task-relevant rep-
resentations in an active state over time without being
overcome by the activity from the stronger processing
pathway.
At a mechanistic level, the model is very simple, in-
volving only a handful of units and relying on the sim-
ple idea that additional activation (from the frontal cor-
tical units) can support processing in a weaker pathway.
Nevertheless, these simple principles are sufficient to
replicate important aspects of the behavioral data. Fur-
thermore, we think that these same simple principles
can be applied to understanding the role of frontal cor-
tex in much more complex tasks (e.g., problem solving).
These more complex tasks would require more sophisti-
cated control mechanisms for activation-based process-
ing, and a richer vocabulary of representations that can
be combined in complex ways — you should recognize
these as the two central problems outlined in the intro-
duction. We begin to address the control problem in
the subsequent simulation, which is followed by a more
detailed discussion of the representation issue. Thus,
the Stroop model establishes the basic principles of top-
down control that are elaborated in the subsequent sec-
tions.
The standard pattern of reaction time performance
in the Stroop task is shown in figure 11.3 (data from